Leo Kaiya, LaVonne Roberts, and Bernard Muller, "A Word That Stays: First-Word Stability in Streaming Dysarthric Speech Recognition in the Speech Accessibility Project Challenge"
Abstract
We describe thirteen of our submissions to SAPC2 Track 2, which ranks streaming recognizers of dysarthric speech on CPU by character error rate and latency. Latency combines the time until the first displayed word stops changing and the delay from the end of the audio to the final transcript; on our four systems in the weekly reports, the first was 75 to 88% of the total. We built three designs. In two-model systems, a small streaming model displays words and a larger model writes the final transcript. The two scored so far were our most accurate on Test1 (9.04% and 9.07% error at total latencies of 1,023 and 835 ms), but the median time to a settled first word ran about 300 ms behind the median time to any first word, a third of their latency. On 21 of 123 development recordings, the final changed or supplied the first word. When the small model also wrote the final on recordings under 8 seconds, such changes there fell from 13 to 1. In fast systems, the first displayed word is locked into the final. In replay, with everything else fixed, the lock raised development error by 1.2 to 2.2 points, and the one fast system scored so far reached 13.28% error on Test1. A stable word is not necessarily a correct one.